Every AI answer is built from sources — the pages the model retrieved (or remembered) and synthesized into a recommendation. Learn which sources those are and you know exactly where your brand needs to show up. Guess wrong and you can publish forever without moving a single answer.
We pulled the citation data from our own daily tracking — 17,000+ citations across two categories we scan every day (B2B SaaS and consumer home goods) — and lined it up against the large public studies. The pattern is remarkably consistent, and it's not where most content budgets are pointed.
Across both categories, citations cluster into four buckets. Corporate sites — vendors, competitors, industry blogs — carry the plurality. But the striking number is how much weight sits with sources brands don't control at all:
Where AI citations come from
Share of 17,000+ citations by source type · two categories tracked daily · Beacon production data, 30 days to Jul 17, 2026
Averaged across our two tracked categories (B2B: 44/32/20/4 · consumer: 54/15/12/19). Half the evidence AI reads lives on pages no brand controls.
Two details worth pausing on. First, the corporate share isn't “your website” — it's mostly other companies'sites: competitors' comparison pages, industry blogs, partner roundups. Your own domain is a minority of even the corporate bucket. Second, the UGC share differs sharply by category — a third of B2B citations vs a sixth of consumer ones in our data — which is why category-specific source mapping beats generic advice every time.
Zoom out to the public mega-studies and the same shape holds. Profound's analysis of 680 million citations found each AI surface has its own favorite sources — ChatGPT leans hard on Wikipedia, Perplexity on Reddit, Google's AI results on YouTube and review platforms.1Semrush's three-month study of 230,000 prompts found Reddit alone among the most-cited domains across AI tools,2 and their analysis of 248,000 AI-cited Reddit posts shows how systematically community threads feed answers.3Ahrefs' 17-million-citation analysis of ChatGPT points the same direction: a concentrated set of trusted domains supplies an outsized share of all evidence.4
Which source types feed which engines
Directional map from the public citation studies + our tracking · 2 = primary, 1 = secondary
No single playbook covers every engine. The engines disagree about sources more than they disagree about brands.
Your website is necessary but nowhere near sufficient.It anchors your claims and feeds retrieval — but it's one voice among dozens the model weighs. Brands that only publish about themselves are structurally invisible to a synthesis machine.
The UGC share is the underpriced asset.Almost a quarter of the citations in our data are community threads — actual people recommending things to each other. Showing up there honestly (as yourself, with disclosure) is the highest-leverage, lowest-competition move in most categories. It's also the least gameable: the FTC now explicitly bans fake reviews and undisclosed astroturfing,5 and models are getting better at weighting authentic threads.
Map your category before you write a word. The 49/24/16/11 split above is an average — your category has its own fingerprint. Pull the actual citations from the answers in your space, list the ten domains that keep appearing, and make those ten your entire earned-media roadmap. Specific beats general by miles here. (That list is exactly what our gap-analysis walkthrough produces.)
That per-category map is what Beacon extracts automatically: every citation behind every AI answer in your space, ranked by how often each domain decides the outcome — so your earned-media roadmap starts as a list of real URLs, not a hunch.
Beacon extracts every citation from the AI answers in your space and ranks the domains that decide who gets recommended.